From my personal POV, and to be perfectly honest, that sounds like a lot of work for little gain. Despite some quirks our current CDR is working well enough, largely doing what we need it to and apparently scaling fine across millions of EHRs.
Another thought that just hit me: How do the compute requirements compare between a workflow using a less performant persist and query method, and one highly dependent on LLMs, both in order to abstract away a lot of technical complexity?